Unified Arab Emirates (UAE) has released an open source model that performs advanced inference and the best offerings from both the US and China.
The new model, K2, comes from researchers at Mohamed Bin Zayed University of Artificial Intelligence (Mbzuai), in Abu Dhabi, the capital of the UAE. This model is one of the first so-called “sovereign” AI models to incorporate the technological advances required for inference, and will be made available free of charge by G42, the Emirati technology conglomerate backed by the Abu Dhabi Sovereign Wealth Fund. The G42 runs the model on a cluster of Cerberas chips, which replaces Nvidia's hardware.
K2 Thinking is one of the UAE's contributions to global races, demonstrating talent in technology that is widely expected to have great economic and geopolitical implications. The US and China are considered dominant players in this contest. However, many small countries, especially those with substantial wealth that invest, are also competing to develop their own “sovereign” AI models.
The K2 is a relatively modest size and has 32 billion parameters. It's not a complete big linguistic model, but a model specialized for reasoning, allowing complex questions to be answered through simulated types of deliberations rather than quickly integrating information to provide output. On such tasks, researchers say it works on par with Openai and Deepseek's inference models, with over 200 billion parameters.
“This is a technological innovation, or, in my opinion, it's confusion,” Eric Singh, president and chief AI researcher at Mbzuai, told WIRED ahead of today's announcement.
According to Xing, the model shows a particularly effective combination of many recent technological innovations. These include tweaking of long strings of simulated inference. Agent planning process that breaks down problems in a variety of ways. Reinforcement learning that trains models to reach the correct answers validated. Other innovations allow models to be delivered very efficiently with Celebras chips.
“It's whether other people want to learn from us, how to make more powerful model features and more powerful features and how to make more powerful features,” Xing said.
Xing adds that he believes the K2 was developed using thousands of GPUs (he refused to give an exact number), and the final training run included a 200-300 chip. The plan is to incorporate K2 into full LLM in the coming months. Mbzuai has opened the model and has published a technical report detailing how different innovations combined to create it.
Other countries in the Middle East, including Saudi Arabia, have also invested heavily in AI infrastructure and research. President Donald Trump traveled to the region in May and announced a number of AI deals involving high-tech American companies.
The UAE leadership has invested billions and established itself as a strategically important research hub. The country has already revealed some cutting-edge AI research and has established a former post base in Silicon Valley. The UAE reduced ties with China in exchange for the access to US silicon needed to train frontier models.
Peng Xiao, CEO of G42 and Mbzuai's board member, said in a statement: “By proving that smaller, more resourceful models rival the biggest systems, this achievement shows how Abu Dhabi is shaping the next wave of global innovation.”
